Hierarchical Meta Alignment for cross-domain object detection

Abstract

Unsupervised domain adaptation (UDA) aims to adapt an object detector from a labeled source domain to an unlabeled target domain. In this task, multiple sub-tasks of different nature are involved, yet existing methods simply sum up the losses and train all the sub-tasks jointly. We, however, find that inconsistent optimization goals between different sub-tasks lead to limited adaptation performance. Specifically, from our analysis, we find notable gradient discrepancies between sub-tasks in a domain adaptive object detector, and especially significant conflicts between domain alignment and detection sub-tasks. Based on this analysis, we propose to solve UDA object detection from a multi-task learning perspective. Specifically, we divide all sub-tasks into two groups, and alleviate both inter-group and intra-group inconsistency via a novel Hierarchical Meta Alignment (HMA) method. At the first level, we construct a Meta Optimization Block (MOB) for each inter-group task pair, which is optimized via the Model-Agnostic Meta-Learning (MAML) algorithm. At the second level, all MOBs are optimized sequentially via the Reptile algorithm. Experimental results on various adaptation scenarios show that our proposed method outperforms previous methods.